“Make New Friends, but Keep the Old” - Recommending People on SN sites Jilin Chen, Werner Geyer, Casey Dugan, Michael Muller, Ido Guy CHI2009 June 1, 2011 Hyewon Lim

Outline  Introduction  Beehive  People Recommendation Algorithms  Experiment I: Personalized Survey  Experiment II: Controlled Field Study  Discussion and Conclusion 2

Introduction  Users connect to both friends they already know offline & new friends they discover on the site –Many users of popular social networking sites primarily communicate with people they already know offline –Users of enterprise social networking sites find valuable contacts not yet known to them, or connecting to weak ties  Finding known contacts and interesting new friends to connect with on the site can both be a challenge 3

Introduction  “People You May Know” on Facebook –Based on a “friend-of-a-friend” approach  Recommending people on social networking sites –Different from traditional recommendations of books, movies, etc., due to the social implication of friending 4

Introduction  Social implications of “friending” –Social dynamics can be obstacles in accepting recommendations  Could be more prominent if unknown or barely known people are recommended –Lack enough motivation  Despite the difficulty… –Connecting with weak ties or unknown but similar people can be more valuable to users than merely re-finding existing strong ties _Granovetter73 5

Beehive  An enterprise social networking site within IBM  Launched in Sep  38,000 users with an average of 8.2 friends per user ( )  Friends are directional –Could be a non-reciprocal friendship  A user can request to be introduced to another user through Beehive 6

People Recommendation Algorithms  Algorithms –Utilize social network structure and based on content similarity  Content Matching  Content-plus-link  Friend-of-a-friend  SONAR 7

People Recommendation Algorithms Algorithm 1: Content Matching  Find users associated with similar content on Beehive  Create a bag-of-words representation of each user –Word vector to describe u: V u = (v u (w 1 ), …, v u (w m ))  v u (w i ) describes the strength of u’s interest in word w i, calculated using a TF-IDF  Similarity of two users –measured by the cosine similarity of V a and V b 8

People Recommendation Algorithms Algorithm 2: Content-plus-Link  Motivation –Disclose a network path to a weak tie or unknown person  Introduce ‘valid social link’ –Computes similarity in the same way as the content matching algo –Boost the similarity by 50% if a valid social link  Favors people in close social network proximity to the user over people more disconnected from the user in the social network 9

People Recommendation Algorithms Algorithm 3: Friend-of-a-Friend  Leverages only social network information of friending –Requires existing friends  Recommendation candidate set RC(u) = { user c | ∃ user a s.t. F(u, a) and F(a, c) }  Mutual friend set MF(u, c) = { user a | F(u, a) and F(a, c) } –Size of MF(u, c) = score of each candidate c for u 10

People Recommendation Algorithms Algorithm 4: SONAR  Based on the SONAR system –Aggregates social relationship information from different public data sources within IBM  organization chart, public DB, patent DB, friending system, people tagging system, project wiki, blogging system –For each data source SONAR computes a normalized relationship score in the range of [0, 1]  SONAR returns a list of users related to u and their aggregated relationship score 11

Experiment I: Personalized Survey  Methodology (with 500 active users) –Overlap ratios between recommendations –Survey 12

Experiment I: Personalized Survey  Result 1: Understanding users’ need –95% of the users considered people recommendations to be useful –Users are interested in connecting to weak ties and meeting new people –Useful information for connecting users to an unknown people 13

Experiment I: Personalized Survey  Result 2: Known vs. unknown, good vs. not good –Result by algorithm 14

Experiment I: Personalized Survey  Result 3: Immediate actions resulted from recommendations –Good recommendations that resulted in actions 15

Experiment II: Controlled Field Study  Methodology (with 3,000 random users) –Divided the 3,000 users randomly into 5 groups  4 groups get recommendations using algorithms each  1 group was a control group that did not get any recommendations –New recommender widget on users Beehive homepage 16

Experiment II: Controlled Field Study  Result 1: Effectiveness of recommender algorithms –Recommendations resulting in connect actions  In contrast to the survey, users rarely chose the introduction option –Users can click a link in widget to view the profile  8% of content matching recommendations, 2.9% for SONAR 17

Experiment II: Controlled Field Study  Result 2: Impact of people recommendations –Immediate goal of recommending people in social networking sites is to increase a user’s network of friends –Compare the number of friends before and after the experiment –People recommendations would impact user activity on the site  Viewed 13.7% more pages vs. 24.4% less pages  Increase in content and comment creation in the exp. Groups (too low) 18

Discussion and Conclusion  In the experiments –Relationship-based algorithms outperform in terms of user response  Relationship-based algorithms –Better at finding known contacts –Perform particularly well for newer users  But, too weak to be meaningful for more established users –FoF can expand their contact list from a few existing contacts –SONAR-like aggregation can take advantage of additional data  Combine the strengths of both types of algorithms –Have an additional benefit of increasing new users’ trust in the system 19